I would add another case (Comparing MATLAB, Python, Python + Numba and R) which I’m not sure why Numba is faster than Julia - Web Appendix for Numerical Language Comparison (Small micro benchmark).
On Slack people used @inbounds and still couldn’t match Numba.
Any idea what gives Numba the edge in those cases?
P. S.
It seems MATLAB is really doing great steps forward,
Probably Apple Libm’s log function. It’s very fast and accurate, and probably exploits that macs only ship with so many different processors. Yes we know it, no we haven’t been able to beat it, and yes it’s super annoying 
I do think there was even a small difference using apple libms log and inbounds .
If I remember correctly someone used Apple’s libm + @inbound and still was a little behind Numba.
So there is a room for improvement regardless of the use of Apple’s library.
I think this issue is related - exp, @fastmath, SVML vectorization.
I hope we’ll see SVML in JuliaPro soon.